Machine learning-based predictions of soil burn severity using spectral and LiDAR remote sensing data

Soil Burn Severity (SBS) indicators qualitatively describe classes of fire-caused changes to soil hydrologic function, as evidenced by soil characteristics and surface fuel and duff consumption. Generally, SBS assessment relies on labor-intensive field measurements, or on the recalibration of remotely sensed Vegetation Burn Severity (VBS) maps. However, our capacity to predict SBS from remote sensing products remains largely unknown. To address this gap, this study explores the potential for predicting SBS directly from remote sensing data using Machine Learning methods (ML). The models developed include the pre-fire and post-fire scenarios, using a total of 29 variables computed from three products for both scenarios. This research offering a novel approach to directly compare pre- and post-fire predictive models of SBS using remote sensing and machine learning. Model calibration and validation was based on 113 SBS field samples collected over 8 wildfire events across an environmental gradient. In the pre-fire model, Random Forest (RF) method achieved higher discriminatory performance (Area Under the Curve (AUC) = 0.90) despite slightly lower overall accuracy (80 %, Kappa = 0.60) compared to XGboost method (81 % accuracy, Kappa = 0.63; AUC = 0.90). The most influential variables in both models included maximum vegetation height and tree canopy cover. For the post-fire model, XGBoost outperformed RF across all metrics, with 77 % accuracy (Kappa =0.54; AUC = 0.76). In both models, the most influential variables included the Relative differenced burn ratio index (Rdnbr) and height skewness. The identification of key variables such as Rdnbr index, skewness, and maximum height LiDAR derived, enhances the practical utility of these models for guiding post-fire mitigation efforts, especially in areas at greater risk of soil degradation and erosion. The present methodology allowed obtaining maps with the predicted SBS occurrence. • XGBoost and Random Forest predict soil burn severity with high accuracy (>80 %). • Pre-fire vegetation height and skewness were key predictors of SBS. • The Relative differenced Normalized Burn Ratio (RdNBR) index was the most important post-fire variable influencing SBS. • Pre-fire predictions of SBS enable proactive identification of erosion-prone areas.